On Sat, 13 Jan 2018, Palmer Dabbelt wrote:
The libm-test-ulps* are just placeholders, so they're almost certainly
nonsense. Aside from those, I started with the aarch64 ABI lists and
You need a real libm-test-ulps (truncate the placeholder and do "make
regen-ulps" for a hard-float configuration).
(It's conceivable that a libm-test-ulps file generated for a hard-float
configuration won't quite suffice for a soft-float one - that another
"make regen-ulps" is needed for soft-float - if any functions have smaller
ulps for hard-float because of use of fma in contracted operations. But
certainly the initial generate needs to be for hard-float so you get ulps
for non-default rounding modes, given a correct math-tests.h.)